Sailfork, timed against GitHub and GitLab.

First over the network with stock git, then on the Linux kernel with only the server's time measured. Every number has its conditions beside it.

Against GitHub and GitLab

Stock git against each host, the way you would run it: a real network between you and the server.

Wall time of stock git against Sailfork, GitHub and GitLab, cloning f-droid/fdroidserver
OperationSailforkGitHubGitLab
List branches and tags264 ms318 ms 1.2x slower616 ms 2.3x slower
Full clone1.56 s2.36 s 1.5x slower2.23 s 1.4x slower
Shallow clone636 ms1.12 s 1.8x slower1.50 s 2.4x slower
Partial clone1.29 s2.17 s 1.7x slower2.64 s 2.1x slower
Stock git 2.55 over the network, from one Mac in San Francisco, on 2026-11-05. The repository is f-droid/fdroidserver (39.9 MiB). Each cell is the median of 5 runs, and times are wall time for the whole command.

The Linux kernel, server time only

This test leaves out the network and HTTP. It measures how long the server takes to send a pack, on one machine, against git 2.54 with a pack bitmap.

OperationServerFirst byteWhole packPack size
Full clone 11,845,633 objectsSailfork140 µs1.89 s2.17 GiB
git 2.544.50 s8.89 s2.03 GiB
Shallow clone 101,713 objectsSailfork180 µs219.2 ms288.16 MiB
git 2.542.86 s13.39 s273.74 MiB
Fetch 100 commits 8,434 objectsSailfork320 µs1.7 ms1.95 MiB
git 2.54195.26 ms216.9 ms9.00 MiB

Server time on linux.git, the median of 5 runs on one Apple M4 Max, with HTTP and the network left out. On a full clone Sailfork sends the whole pack 4.7x faster; on a shallow clone, 61.1x; on a fetch of the last 100 commits, 127.6x. First byte is how long the server works before it sends anything: stock git counts objects and builds a pack first, and Sailfork builds the pack when you push.

The cost is a larger pack: +7.0% on a full clone, because Sailfork sends packs as stored instead of recompressing them for each clone. Serving a full clone also takes 6.86x less CPU per GB sent: 0.591 s against 4.052 s.

Watch it replayed

Pack size, against stock git

  • Full clone
    2.17 GiB
    2.03 GiB
    +7.0%
  • Shallow clone
    288.16 MiB
    273.74 MiB
    +5.3%
  • Fetch 100 commits
    1.95 MiB
    9.00 MiB
    −78.3%
Each row is scaled to its larger pack. Clone packs are a little larger because Sailfork sends them as stored instead of recompressing them for each clone; the fetch pack is smaller.

Caissons and the command line

Per-call times for an agent working through a caisson, and for the sailfork command.

One caisson call

100 µs
Open a caisson
1 µs
Read a file
10 µs
Write a file
500 µs
Commit

Server time per call on sailfork.com, as an agent sees it through the API. A caisson is an agent's own branch and its unsaved edits, kept on Sailfork; nothing is cloned and no machine starts.

The sailfork command

  • sailfork repo view, on sailfork.com 118.6 ms
  • curl, the same request 135.4 ms
  • gh repo view, on github.com 514.7 ms
  • sailfork --version 3.4 ms

Mean of 50 runs with hyperfine on an Apple silicon Mac in San Francisco, November 2026. gh talks to github.com, so its time includes a different network path. The binary is 7.28 MB on macOS.

How the Linux test was measured

one Apple M4 Max, so the network plays no part

linux.git

11,845,633 objects

Sailfork

Sends stored pack segments

git upload-pack 2.54

With a pack bitmap, its fastest setup

In-process clock

First pack byte and whole pack. HTTP and the network are left out.

One warm-up, then 5 runs sorted by time. The middle one, the median, is the number reported.

One repository
A copy of linux.git, the Linux kernel, with 11,845,633 objects.
One machine
Both servers run on one Apple M4 Max, so the network plays no part. Stock git is upload-pack 2.54 with a pack bitmap, its fastest setup.
Server time only
The clock measures how long the server takes to send its first pack byte and the whole pack, in process, so HTTP and the network are left out.
Median of five
Each operation runs 5 times after one warm-up, and the median is reported.

What it leaves out

Measured here

The server sends the pack

1.89 s on Sailfork, first byte at 140 µs

Not measured

The network carries it

Depends on your connection

Not measured

Your machine resolves deltas

Depends on your machine

The three legs of a real clone, not drawn to scale. A full Linux clone on your laptop takes longer than the server’s part, on Sailfork or anywhere else; what the numbers show is that Sailfork starts sending at once.

Loading Sailfork